Anthropic has quietly proposed a draft bill to US lawmakers that would establish new safety rules for advanced artificial intelligence systems. It included limits on when the government could halt their release or use, Politico reported on October 16.
The proposal would allow the government to intervene only when an AI model poses a foreseeable and imminent threat involving more than USD 1 billion in damages or the death or serious injury of more than 50 people, according to the report.
The proposed restrictions would apply to risks linked to biological weapons, AI models escaping their developers’ controls or cyberattacks, Politico reported. The framework would also require developers of the most powerful AI systems to conduct safety tests, disclose findings and obtain independent assessments.
An Anthropic spokesperson told Reuters that the proposal reflects measures outlined in the company's Advanced AI Framework, released earlier this year. The spokesperson said the bill would give authorities the ‘ability to block or deter dangerous deployments of the most powerful models,’ while requiring developers to test their systems and undergo independent evaluations.
The proposal would also establish a formal process for assessing potential threats. Independent evaluators would register with the US Department of Commerce and periodically examine AI developers and their frontier models.
AI companies could pay the independent evaluators, provided their fees were not linked to the outcome of an assessment, Politico reported. Evaluators who identify an ‘imminent enumerated risk’ would have to notify the Commerce Department within 72 hours of completing their review.
The department could then ask a judge to suspend a model if it determined that the system posed a qualifying risk. The draft would specifically bar other US government departments, agencies and officials from independently suspending AI development, deployment or internal use. A federal judge in the District of Columbia could order a suspension, according to the report.
The US Department of Commerce did not immediately respond to Reuters' request for comment. The proposal outlines Anthropic's preferred approach to AI oversight, though the report did not establish whether lawmakers would adopt the draft or advance it into legislation.
Also Read: Anthropic Opens Advanced Claude Models to More Cybersecurity Firms